Differentiable latent structure discovery for interpretable forecasting in clinical time series

2026-04-30

AI Summary

This work addresses the challenges of limited interpretability and difficulty in uncertainty modeling inherent in irregular electronic health records by proposing StructGP and its low-rank extension, LP-StructGP. The method uniquely integrates differentiable structure learning with continuous-time multi-task Gaussian processes, leveraging process convolution, sparse acyclicity constraints, and a softmax gating mechanism to jointly infer a sparse directed acyclic graph among variables and shared latent pathways across patients. Evaluated on MIMIC-IV and PhysioNet, the model significantly outperforms baseline approaches, achieving RMSE as low as 0.58–0.69 and predictive coverage up to 0.96. Notably, structural recovery accuracy approaches near-perfect levels with increasing cohort size, while maintaining well-calibrated uncertainty estimates and strong interpretability.

Abstract

Background: Timely, uncertainty-aware forecasting from irregular electronic health records (EHR) can support critical-care decisions, yet most approaches either impute to a grid or sacrifice interpretability. We introduce StructGP, a continuous-time multi-task Gaussian process that couples process convolutions with differentiable structure learning to uncover a sparse, ordered directed acyclic graph (DAG) of inter-variable dependencies while preserving principled uncertainty. We further propose LP-StructGP, which augments StructGP with latent pathways-shared, temporally shifted trajectories inferred via subject-specific coupling filters and a softmax gating mechanism-to capture cross-patient progression patterns. Both models are trained under sparsity and acyclicity constraints (augmented Lagrangian, Adam) using scalable low-rank updates. Results: In simulations, the approach reliably recovers ground-truth graphs (Structural Hamming Distance approaching 0 as cohorts grow) and pathway assignments (high Adjusted Rand Index). On a MIMIC-IV septic shock cohort (n=1,008; norepinephrine, creatinine, mean arterial pressure), StructGP improves short-horizon (6 h) forecasting over independent-task baselines (average RMSE 0.68 [95%CI: 0.63--0.74] vs. 0.88 [0.83-0.94]) and, with 15 additional inputs, markedly outperforms unstructured kernels (0.63 [0.58-0.69] vs. 3.02 [2.85-3.18]) with superior calibration (coverage 0.96 vs. 0.84). On the PhysioNet Challenge (12k patients, 41 variables), StructGP attains competitive accuracy (MAE 3.72e-2) relative to a state-of-the-art graph neural model while maintaining calibrated uncertainty. Conclusion: These results show that structured process convolutions with latent pathways deliver interpretable, scalable, and well-calibrated forecasting for irregular clinical time series.

Problem

• interpretable forecasting
• clinical time series
• irregular EHR
• uncertainty-aware prediction
• structured dependencies

Innovation

• differentiable structure learning
• structured Gaussian process
• latent pathways
• directed acyclic graph (DAG)
• irregular time series forecasting